摘要
针对决策树(DT)模型缺乏概率背景这一问题,将贝叶斯推理引入DT模型,提出了一种基于贝叶斯推理的决策树(BDT)模型.在假定所含待定参量的先验与似然的前提下,借助贝叶斯推理获得参量的后验,然后运用逆跳马尔科夫链蒙特卡洛算法对后验抽样,最终求出样本属于某一类别的置信度,从而避免了武断判决.BDT模型以抽样代替拆分与剪枝操作,既直观又灵活,同时在抽样时考虑了不同的树结构与递归分割方案,使得分类准确率得以提高.仿真实验结果表明,BDT模型的平均分类准确率与DT模型相比提高了1.7%~3.5%.
Focusing on the problem that conventional decision tree (DT) model lacks of a probabilistic background, the Bayesian inference was introduced into DT, and thus a decision tree (BDT) model based on Bayesian inference was proposed. Under the premise that the prior and likehood of contained parameters needed to be determined has been assumed, the posterior of parameters are obtained through Bayesian inference. Then the posterior is sampled by using reversible jump Markov chain Monte Carlo algorithm, and finally the confidence level of the samples belonged to certain class is solved to avoid any arbitrary decision. In BDT model, the splitting and pruning is substituted by sampling, both are intuitive and flexible, and different tree structures and recurperimental results show that the average classification accuracy is improved by 1.7 % -3.5 % compared to DT model.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2006年第8期888-891,共4页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(90207012)
关键词
决策树
贝叶斯推理
逆跳马尔科夫链蒙特卡洛
分类准确率
递归分割
decision tree; Bayesian inference; reversible jump Markov chain Monte Carlo; classification accuracy rate;recursive partition